CohEx: A Generalized Framework for Cohort Explanation

Fanyu Meng, Xin Liu, Zhaodan Kong, Xin ChenAAAI 2025 · DOI

Introduction

In the field of eXplainable AI (XAI), researchers typically navigate a trade-off between two extremes. Global explanations offer a holistic view of model behavior across an entire dataset but often suffer from "aggregation bias," failing to capture how a model treats specific subgroups. Conversely, local explanations (such as LIME or SHAP) provide high precision for individual predictions but are often overwhelming to interpret at scale and lack robustness against minor input perturbations.

CohEx (Cohort Explanation) introduces a middle ground. It is a generalized framework designed to identify specific subgroups, or cohorts, and generate concise, robust explanations tailored to model behavior within those specific regions.

The CohEx Framework

CohEx converts standard local feature importance methods into cohort-level insights through an iterative feedback loop. Unlike previous methods that simply average pre-calculated local explanations, CohEx ensures "Cohort Locality" by recomputing importance scores using only the samples within a specific cohort. This prevents information from outside the region from biasing the explanation.

Example of a cohort-level feature importance explanation generated by CohEx, highlighting the relative impact of Age versus Family History (FH).

The framework follows a four-step process:

  1. Initialization: Data is partitioned into $k$ initial cohorts.
  2. Localized Recomputation: Local importance scores (e.g., DeepLIFT, SHAP) are recomputed using only cohort-specific data.
  3. Supervised Clustering: A supervised clustering algorithm (SRIDHCR) refines cohort boundaries by considering both raw features and the newly computed importance scores.
  4. Iteration: The process repeats until the clustering loss stabilizes, ensuring that cohorts are optimized for both feature similarity and explanation consistency.

Experimental Validation

The authors evaluated CohEx using synthetic patient data, a real-world bike-sharing regression task, and the MNIST vision dataset. A primary challenge in these tasks is capturing non-linear decision boundaries that a global model might oversimplify.

Synthetic patient classification task showing the discrepancy between axis-aligned tree boundaries and the true optimal decision boundary.

In the synthetic patient classification task (Age vs. Family History), CohEx demonstrated its ability to identify "hidden" patterns. While global trends suggested Age was the primary driver of predictions, CohEx successfully isolated a specific cohort where Family History was the dominant factor—a nuance missed by baseline methods like VINE and REPID.

Key Insights and Results

  • Locality and Faithfulness: CohEx yielded significantly lower "locality loss" compared to baselines, meaning its explanations were more faithful to the specific regions they described.
  • Stability: The framework demonstrated higher stability, with importance scores remaining consistent even when small amounts of new data were introduced to a cohort.
  • Interpretability Spectrum: By adjusting the number of cohorts ($k$), users can find a "sweet spot" between a single global overview and thousands of individual local explanations.

Limitations and Future Directions

While CohEx provides a robust bridge between local and global XAI, the iterative recomputation of importance scores can be computationally intensive for complex deep learning models. Future work will focus on improving the stability of cohort boundaries and exploring the use of human-readable decision rules to define these regions more intuitively.

Summary updated

BibTeX
@inproceedings{meng2025cohex,
  author = {Fanyu Meng and Xin Liu and Zhaodan Kong and Xin Chen},
  title = {{CohEx: A Generalized Framework for Cohort Explanation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34140},
}
Figure